FedFM: A federated few-shot learning method by comparison network and model calibration

Federated Learning (FL) is a flexible and efficient approach for leveraging distributed data through parameter upload and aggregation. However, the practical applicability of current FL methodologies is constrained by issues such as label scarcity and client heterogeneity. (1) Label scarcity arises from users lacking the capability or willingness to annotate local data. This constraint results in inadequately trained models and obstructs the inference phase in FL. (2) Client heterogeneity stems from the asynchronous update mechanism of shared models. The server calculates a weighted average of the parameters uploaded by participating clients, overlooking the training variations and model-calibration differences among them. Herein, we present a novel federated few-shot learning framework called federated few-shot and model-calibration to address the above challenges. Initially, we use strong data augmentation to improve feature extraction efficiency and train the comparison network that maps unlabeled data into labels of a labeled support set, eliminating the necessity for fine-tuning to accommodate new class types. In addition, the inherent heterogeneity of data across multiple clients poses a challenge that frequently results in performance deterioration. To address this issue, we propose a model-calibration update method that aggregates model parameters weighted by the expected calibration errors obtained using local models. Compared with existing FL approaches, our method achieves improved one-shot classification accuracy from 85.42% to 92.61% and from 42.85% to 47.92% on the Omniglot and Mini-ImageNet datasets, respectively. In addition to providing improved performance on few-shot learning and privacy preservation, our method can be easily extended to zero-shot learning.